Enhancing Federated Learning Performance for IoT Anomaly Detection Under Label-Skewed Data

Federated Learning (FL) is a promising approach for IoT anomaly detection due to its privacy-preserving nature. However, it is common in an IoT scenario for the devices to carry different types of data. FL algorithms that use aggregation algorithms like FedAvg tend to experience a decline in performance under this type of skewed data distribution. Therefore, it becomes important to deal with the problem of data heterogeneity to improve FL performance in terms of precisely detecting anomalies over normal data. To address the label-skewness in FL in general, a previous study proposed a data sharing technique with the aim of introducing data homogeneity in the local data of the clients involved in a FL setup. In this paper, some modifications are made to this technique to reduce the associated client data privacy risks and evaluate its efficacy in an IoT environment by conducting experiments on a real-world IoT dataset (X-IIoTID). The results demonstrate that this approach shows significant improvements under label-skewed data and its ease of implementation has the potential to make it a promising solution for resource-constrained domains like IoT anomaly detection.